Deep Learning-Based Classification and Segmentation of Sperm Head and Flagellum for Image-Based Flow Cytometry

Abstract

Image-Based Flow Cytometry (IBFC) is a potent tool for the detailed analysis and quantification of cells in intricate samples, facilitating a comprehensive understanding of biological processes. This study leverages the ResNet50 model to address IBFC’s object-defocusing issue, an inherent challenge when imaging a 3D object with stationary optics. A dataset of 604 mouse sperm IBFC images (both bright field and fluorescence) underpins the exceptional capability of the ResNet50 model to reliably identify optimally focused images of the sperm head and flagella (F1-Score of 0.99). A U-Net model was subsequently employed to accurately segment the sperm head and flagellum in images selected by ResNet50. Notably, the flagellum presents a significant challenge due to its sub-diffraction transversal dimensions of 0.4 to 1 micrometers, resulting in minimal light intensity gradients. The U-Net model, however, demonstrates exceptional efficacy in precisely segmenting the flagellum and head (dice = 0.81). The combined ResNet50/U-Net approach offers significant promise for enhancing the efficiency and reliability of sperm analysis via IBFC, and could potentially drive advancements in reproductive research and clinical applications. Additionally, these innovative strategies may be adaptable to the analysis of other cell types.